boore_atkinson_2008_ground motion prediction equations for the average horizontal component of pga,...

8/9/2019 BOORE_ATKINSON_2008_Ground Motion Prediction Equations for the Average Horizontal Component of PGA, PGV, A

1/40

Ground-Motion Prediction Equationsfor the Average Horizontal Componentof PGA, PGV, and 5%-Damped PSA

at Spectral Periods between 0.01 sand10.0 s

David M. Boorea)

and Gail M. Atkinson,b)

M.EERI

This paper contains ground-motion prediction equations (GMPEs) foraverage horizontal-component ground motions as a function of earthquake

magnitude, distance from source to site, local average shear-wave velocity, andfault type. Our equations are for peak ground acceleration (PGA), peak ground

velocity (PGV), and 5%-damped pseudo-absolute-acceleration spectra (PSA)at periods between0.01 sand10 s. They were derived by empirical regressionof an extensive strong-motion database compiled by the PEER NGA (Pacific

Earthquake Engineering Research Centers Next Generation Attenuation)

project. For periods less than 1 s, the analysis used 1,574 records from 58mainshocks in the distance range from 0 km to 400 km (the number ofavailable data decreased as period increased). The primary predictor variables

are moment magnitude M, closest horizontal distance to the surfaceprojection of the fault plane RJB, and the time-averaged shear-wave velocityfrom the surface to 30 m VS30. The equations are applicable forM = 5 8,

RJB200 km, andVS30=1801300 m/ s. DOI: 10.1193/1.2830434

INTRODUCTION

Ground-motion prediction equations (GMPEs), giving ground-motion intensity mea-sures such as peak ground motions or response spectra as a function of earthquake mag-nitude and distance, are important tools in the analysis of seismic hazard. These equa-tions are typically developed empirically by a regression of recorded strong-motionamplitude data versus magnitude, distance, and possibly other predictive variables. The

equations in this report were derived as part of the Pacific Earthquake Engineering Re-search Centers Next Generation Attenuation project (PEER NGA; Power et al. 2008),using an extensive database of thousands of records, compiled from shallow crustalearthquakes in active tectonic environments worldwide. These equations represent a sub-stantial update to GMPEs that were published by Boore and his colleagues in 1997(Boore et al. 1997, hereafter BJF97; note that BJF97 summarized work previously

published by Boore et al. in 1993 and 1994). The 1997 GMPEs of Boore et al. werebased on a fairly limited set of data in comparison to the results of this study. The in-

a)U.S. Geological Survey, MS 977, 345 Middlefield Rd., Menlo Park, CA 94025

b)Department of Earth Sciences, University of Western Ontario, London, Ont. Canada N6A 5B7

99

Earthquake Spectra, Volume 24, No. 1, pages 99138, February 2008; 2008, Earthquake Engineering Research Institute


2/40

crease in data quantity, by a factor of approximately 14, is particularly important forPSA; in addition, PGV equations are provided in this study (but were not given in

BJF97). The amount of data used in regression analysis is an important issue as it bearsheavily on the reliability of the results, especially in magnitude and distance ranges thatare important for seismic hazard analysis.

This paper is a condensation of our final project report published by PEER (Booreand Atkinson 2007); the reader may refer to that document for more details and a num-

ber of relevant appendices. We will refer to that report as BA07.

DATA

DATA SOURCES

The source of the strong ground-motion data for the development of the GMPEs inthis study is the database compiled in the PEER NGA project (Chiou et al. 2008); the

aim of that project was to develop empirical GMPEs using several investigative teams toallow a range of interpretations (this paper is the report of one team). The use of thisdatabase, referred to as the NGA Flatfile, was one of the ground rules of the GMPE

development exercise. However, investigators were free to decide whether to use the en-tire NGA Flatfile database, or to restrict their analyses to selected subsets.

In addition to the data in the NGA Flatfile, we also used data compiled by J. Boat-wright and L. Seekins for three small events, and data from the 2004 Parkfield, Califor-nia, mainshock from the Berkeley Digital Seismic Network station near Parkfield, aswell as data from the Strong-Motion Instrumentation Program of the California Geo-logical Survey and the National Strong-Motion Program of the United States Geological

Survey. These additional data were used in a study of the distance attenuation functionthat constrained certain regression coefficients, as discussed later, but were not included

as part of the final regression (to be consistent with the NGA ground rules regardingthe database for regression).

RESPONSE VARIABLES

The ground-motion parameters that are the dependent variables of the GMPEs (alsocalled response variables or ground-motion intensity measures) include peak ground ac-celeration (PGA), peak ground velocity (PGV), and response spectra (PSA, the 5%-damped pseudo-acceleration), all for the horizontal component. In this study, the re-sponse variables are not the simple geometric mean of the two horizontal component (as

was used in BJF97), but rather are measures of geometric mean not dependent on theparticular orientation of the instruments used to record the horizontal motion. The mea-sure used was introduced by Boore et al. (2006). In that paper a number of orientation-

independent measures of ground motion were defined. In this report we use GMRotI50(which we abbreviate GMRotI); this is the geometric mean determined from the 50th-

percentile values of the geometric means computed for all non-redundant rotation angles

and all periods less than the maximum useable period. The advantage of using anorientation-independent measure of the horizontal component amplitude can be appre-ciated by considering the case in which the motion is perfectly polarized along one com-

100 D. M. BOORE AND G. M. ATKINSON


3/40

ponent direction; in this case the geometric mean would be 0. In most cases, however,the differences between the geometric mean and GMRotI are not large, so that our re-

sponse variable can be thought of in simple terms as an average horizontal component.This paper includes GMPEs for PGA, PGV, and 5%-damped PSA for periods be-

tween0.01 sand10 s. Equations for peak ground displacement (PGD) are not included.In our view, PGD is too sensitive to the low-cut filters used in the data processing to be

a stable measure of ground shaking. In addition there is some bias in the PGD valuesobtained in the NGA data set from records for which the low-cut filtering was not per-formed as part of the NGA project. Appendix C in BA07 contains a short discussion ofthese points. We recommend using response spectra at long periods instead of PGD.

Data were excluded from our analysis based on a number of criteria, the most im-portant of which (in terms of number of records excluded from the analysis) is that no

recordings from obvious aftershocks were used. Aftershock records were not used be-cause of some concern that the spectral scaling of aftershocks differs from mainshocks(see Boore and Atkinson 1989 and Atkinson 1993). This restriction cut the data set al-most in half because a substantial number of the records in the NGA Flatfile are after-shocks of the 1999 Chi-Chi earthquake. The other exclusion criteria that were appliedare listed in Table 2.1 of BA07. Response variables were excluded for oscillator periods

greater than TMAX (the inverse of the lowest useable frequency entry in the NGAFlatfile).

We did not use singly recorded earthquakes. Table 1 lists all earthquakes used in ourdata analysis, along with the number of stations used per earthquake (for an oscillator

period of0.2 s).

A potential bias in regression results can result from not including low-amplitudedata from distance ranges for which larger amplitude data for the same earthquake are

included in the data set. There are several reasons that low-amplitude data might not beincluded: it can be below trigger thresholds of instruments, which will cause the record-ing to begin at some point during the S-wave arrival, it can be too small to digitize, or itcan be below the noise threshold used in determining low-cut filter frequencies. Any col-lection of data in a small distance range will have a range of amplitudes because of the

natural variability in the ground motion (due to such things as source, path, and site vari-ability). At distances far enough from the source (depending on magnitude), some of thevalues in the collection will be below the amplitude cutoff and would therefore be ex-cluded. If only the larger motions (above the amplitude cutoff) were included, this wouldlead to a bias in the predicted distance decay of the ground motionthere would be atendency for the predicted ground motions to decay less rapidly with distance than the

real data. BJF97 attempted to avoid this bias by excluding data for each earthquake be-yond the closest distance to an operational, non-triggered station (most of the data used

by BJF97 were obtained on triggered analog stations). Unfortunately, information is notavailable in the NGA Flatfile that would allow us to apply a similar distance cutoff, atleast for the case of triggered analog recordings. Furthermore, a similar bias might also

exist in digital recordings because of the presence of long-period noise that is indepen-

GROUND-MOTION PREDICTION EQUATIONS FOR AVERAGE HORIZONTAL COMPONENT OF PGA, PGV, & PSA 101


4/40

Table 1. Events used in analysis, for a period of 0.2 s, giving type of earthquake (S=strike-slip,

N = normal, R = reverse), number of observations (NOBS), range ofRJB in km, and NGA Flat-

file event identification number

NAME YEAR MODY M DIP

DEPTH

(km) TYPE NOBS

RJB

RANGE EQID

Parkfield 1966 0628 6.19 90 10 S 4 1018 25

Borrego Mtn 1968 0409 6.63 78 8 S 2 129222 28

San Fernando 1971 0209 6.61 50 13 R 31 14218 30

Hollister-03 1974 1128 5.14 90 6 S 2 910 34

Friuli, Italy-01 1976 0506 6.50 12 5 R 5 15102 40

Tabas, Iran 1978 0916 7.35 25 6 R 7 0194 46

St Elias, Alaska 1979 0228 7.54 12 16 R 2 2680 142

Coyote Lake 1979 0806 5.74 80 10 S 7 034 48

Norcia, Italy 1979 0919 5.90 64 6 N 3 231 49

Imperial Valley-06 1979 1015 6.53 80 10 S 33 049 50

Livermore-01 1980 0124 5.80 85 12 S 5 1553 53

Anza (Horse Canyon)-01 1980 0225 5.19 70 14 S 5 639 55

Mammoth Lakes-01 1980 0525 6.06 50 9 N 2 15 56

Victoria, Mexico 1980 0609 6.33 90 11 S 4 639 64

Irpinia, Italy-01 1980 1123 6.90 60 10 N 12 760 68

Westmorland 1981 0426 5.90 90 2 S 6 619 73

Coalinga-01 1983 0502 6.36 30 5 R 44 2455 76

Borah Peak, ID-01 1983 1028 6.88 52 16 N 2 8385 87

Morgan Hill 1984 0424 6.19 90 9 S 24 371 90

Lazio-Abruzzo, Italy 1984 0507 5.80 48 14 N 5 1349 91

Hollister-04 1986 0126 5.45 90 9 S 3 1113 98

N Palm Springs 1986 0708 6.06 46 11 R 30 078 101Chalfant Valley-01 1986 0720 5.77 90 7 S 5 624 102

Chalfant Valley-02 1986 0721 6.19 55 10 S 10 651 103

San Salvador 1986 1010 5.80 85 11 S 2 24 108

Whittier Narrows-01 1987 1001 5.99 30 15 R 106 082 113

Superstition Hills-02 1987 1124 6.54 90 9 S 11 127 116

Loma Prieta 1989 1018 6.93 70 18 R 73 0117 118

Upland 1990 0228 5.63 77 5 S 3 772 143

Manjil, Iran 1990 0620 7.37 88 19 S 7 13175 144

Sierra Madre 1991 0628 5.61 50 12 R 8 346 145

Roermond, Netherlands 1992 0413 5.30 68 15 N 3 55101 122

Cape Mendocino 1992 0425 7.01 14 10 R 6 040 123

Landers 1992 0628 7.28 90 7 S 68 2190 125

Big Bear-01 1992 0628 6.46 85 13 S 39 7147 126Little Skull Mtn, NV 1992 0629 5.65 70 12 N 8 1499 152

Northridge-01 1994 0117 6.69 40 18 R 154 0148 127

Kobe, Japan 1995 0116 6.90 85 18 S 12 0158 129

Kozani, Greece-01 1995 0513 6.40 43 13 N 3 1479 130

Dinar, Turkey 1995 1001 6.40 45 5 N 4 0255 134



5/40

dent of the distance from the source to the station. Consequently, the obtained distance

dependence for small earthquakes and long periods may be biased towards a decay that

is less rapid than the true decay.

PREDICTOR VARIABLES

The primary predictor variables (independent variables in the regression analysis) are

moment magnitude M, RJB distance (closest distance to the surface projection of thefault plane), andVS30for site characterization. (VS30is the time-averaged shear-wave ve-locity over the top 30 m, calculated as the inverse of the average shear-wave slownessfrom the surface to a depth of30 m[although slowness is simply the inverse of velocity,it has a number of useful properties, as discussed in Boore and Thompson 2007].) The

RJBdistances estimated by R. Youngs, as described in Appendix B of Chiou and Youngs(2006), were used for earthquakes with unknown fault geometry.

We also considered the effect of fault type (i.e., normal, strike-slip, and reverse). The

fault type was specified by the plunge of the P- andT-axes, as shown in the legend toFigure 1 (Appendix D of BA07 contains a more complete description). While there are

some advantages to using P- andT-axes, Figure 1 shows that the simple classificationused by BJF97 (rakes angles within 30 of horizontal are strike-slip, angles from 30 to

Table 1. (cont.)

NAME YEAR MODY M DIP DEPTH(km) TYPE NOBS RJBRANGE EQID

Northwest China-01 1997 0405 5.90 68 23 S 2 1249 153

Northwest China-02 1997 0406 5.93 30 31 N 2 2037 154

Northwest China-04 1997 0415 5.80 43 22 N 2 2135 156

Kocaeli, Turkey 1999 0817 7.51 88 15 S 26 1316 136

Chi-Chi, Taiwan 1999 0920 7.62 30 7 R 380 0169 137

Hector Mine 1999 1016 7.13 77 5 S 82 10233 158

Dzce, Turkey 1999 1112 7.14 65 10 S 22 0188 138

Yountville 2000 0903 5.00 90 10 S 24 894 160

Big Bear-02 2001 0210 4.53 90 9 S 41 2292 161

Mohawk Val, Portola 2001 0810 5.17 81 4 S 6 67126 162

Anza-02 2001 1031 4.92 78 15 S 72 10133 163

Gulf of California 2001 1208 5.70 59 10 S 11 72130 164

CA/Baja Border Area 2002 0222 5.31 74 7 S 9 4097 165

Gilroy 2002 0514 4.90 84 10 S 34 2130 166

Yorba Linda 2002 0903 4.27 88 7 S 12 636 167

Nenana Mountain,

Alaska

2002 1023 6.70 90 4 S 33 105280 168

Denali, Alaska 2002 1103 7.90 71 5 S 23 0276 169

Big Bear City 2003 0222 4.92 72 6 S 33 24146 170



6/40

150 are reverse, and angles from30 to 150 are normal) gives essentially the sameclassifications as obtained using P- andT-axes, at least for the data set we used.

All of the predictor variables were taken from the NGA database.

DISTRIBUTION OF DATA BY M, RJB, FAULT TYPE,AND SITE CLASS

Representative M andRJB distributions of the data used in developing our GMPEsare shown in Figure 2, with the symbols representing different fault types. The totalnumber of recordings for the analysis (after all exclusions) is 1,574 for periods out to

1 s, with a slight decrease at 2 s, and a rapid fall off in the number of available data atperiods longer than 2 s. The distributions of the data over the predictor variable spacenecessarily influence the GMPEs. Note in particular the lack of data at close distancesfor small earthquakes. This means that the near-source ground motions for small eventswill not be constrained by observations. For long oscillator periods, there are very few

data for small earthquakes at any distance (the points in Figure 2 for M = 5 and T

Figure 1. Distribution of the data we used in rake-angle and dip-angle space. The horizontal

gray lines indicate boundaries between fault types used by BJF97, and the symbols and colors

indicate our classification based on the plunges of the P- and T-axes (our classification schemeis indicated in the legend; see Appendix D in BA07).



7/40

=10 s are all from a single eventthe 2000 Yountville, California earthquake), so themagnitude scaling at long periods will be poorly determined for small magnitudes.

The widest range of magnitudes is for strike-slip earthquakes (4.37.9), while the

narrowest range is for normal-slip earthquakes (5.36.9). This suggests that the magni-tude scaling is better determined for strike-slip than for normal-slip earthquakesa

problem that we circumvented by using a common magnitude scaling for all types ofevents, as discussed later.

The bulk of the data are from class C and D sites, which range from soft rock to firmsoil; very few data were from class A sites (hard rock). More detail can be found in

Appendix A, which includes two possible sets ofVS30 values to use in evaluating ourequations for a particular NEHRP site class.

THE EQUATIONS

Following the philosophy of Boore et al. (1993, 1994, 1997), we seek simple func-tional forms for our GMPEs, with the minimum required number of predictor variables.

We started with the simplest reasonable form for the equations (that used in BJF97), andthen added complexity as demanded by comparisons of the predictions of ground mo-tions from the simplest equations with the observed ground motions. The selection offunctional form was heavily guided by subjective inspection of nonparametric plots of

data; many such plots were produced and studied before commencing the regressionanalysis. For example, the BJF97 equations modeled the far-source attenuation of am-

Figure 2. Distribution of data used to derive our regression equations for PGA and for PSA at

a period 10.0 s, differentiated by fault type (points withRJBless than 0.1 km plotted at 0.1 km).

The overall distributions for periods less than about 4 s are similar to those for PGA, although

there are fewer recordings (the number of available recordings decreases noticeably for periods

longer than 2 s).



8/40

plitudes with distance by a simple function that had no magnitude dependence and nocurvature at greater distances. This form appeared sufficient for the maximum distance

range of80 km specified for the BJF97 GMPEs. The data, however, clearly show thatcurvature of the line is required to accommodate the effects of anelastic attenuation

when modeling data beyond80 km; furthermore, the data show that the effective geo-metric spreading factor is dependent on magnitude. To accommodate these trends, we

(1) added an anelastic coefficient to the form of the equations, in which lnY is pro-portional to R (where Y is the response variable), and (2) introduced a magnitude-dependent geometrical spreading term, in which lnY is proportional to lnR and the

proportionality factor is a function ofM. These features allow the equations to predict

amplitudes to 400 km; the larger size of the NGA database at greater distances and forlarger magnitudes, in comparison to that available to BJF97, enabled robust determina-tion of the additional coefficients. Our functional form does not include such factors asdepth-to-top of rupture, hanging wall/footwall terms, or basin depth, because residual

analysis does not clearly show that the introduction of such factors would improve theirpredictive capabilities on average. The equations are data-driven and make little use ofsimulations. They include only those terms that are truly required to adequately fit theobservational database, according to our analysis. Our equations may provide a usefulalternative to the more complicated equations provided by other NGA models, as theywill be easier to implement in many applications.

Our equation for predicting ground motions is:

lnY=FMM+FDRJB,M+FSVS30,RJB,M+ T, 1

In this equation, FM,FD, andFSrepresent the magnitude scaling, distance function, andsite amplification, respectively. M is moment magnitude, RJB is the Joyner-Boore dis-tance (defined as the closest distance to the surface projection of the fault, which is ap-

proximately equal to the epicentral distance for events ofM6), and the velocity VS30is the inverse of the average shear-wave slowness from the surface to a depth of30 m.The predictive variables are M, RJB, andVS30; the fault type is an optional predictivevariable that enters into the magnitude scaling term as shown in Equation 5a and 5b be-

low. is the fractional number of standard deviations of a single predicted value oflnYaway from the mean value of lnY (e.g., =1.5 would be 1.5 standard deviationssmaller than the mean value). All terms, including the coefficient T, are period depen-

dent. Tis computed using the equation

T=2 + 2, 2

where is the intra-event aleatory uncertainty and is the inter-event aleatory uncer-

tainty (this uncertainty is slightly different for cases where fault type is specified andwhere it is not specified; we distinguish these cases by including a subscript on ).



9/40

THE DISTANCE AND MAGNITUDE FUNCTIONS

The distance function is given by:

FDRJB,M=c1+c2M MreflnR/Rref+c3RRref, 3

where

R=RJB2 +h2 4

andc 1, c 2, c3, Mref, Rref, andh are the coefficients to be determined in the analysis.

The magnitude scaling is given by:

a) MMh

FMM=e1U+e2SS+e3NS+e4RS+e5M Mh+e6M Mh2, 5a

b) MMh

FMM=e1U+e2SS+e3NS+e4RS+e7M Mh , 5bwhere U, SS, NS, andRS are dummy variables used to denote unspecified, strike-slip,normal-slip, and reverse-slip fault type, respectively, as given by the values in Table 2,

andMh, the hinge magnitude for the shape of the magnitude scaling, is a coefficient tobe set during the analysis.

SITE AMPLIFICATION FUNCTION

The site amplification equation is given by:

FS=FLIN+FNL, 6

where FLINandFNL are the linear and nonlinear terms, respectively.

The linear term is given by:

FLIN=blinlnVS30/Vref, 7

where blin is a period-dependent coefficient, andVref is the specified reference velocity=760 m/ s, corresponding to NEHRP B/C boundary site conditions; these coefficients

Table 2. Values of dummy variables for different

fault types

Fault Type U SS NS RS

Unspecified 1 0 0 0

Strike-slip 0 1 0 0

Normal 0 0 1 0

Thrust/reverse 0 0 0 1



10/40

were prescribed based on the work of Choi and Stewart (2005; hereafter CS05); theyare empirically based but were not determined by the regression analysis in our study.

The nonlinear term is given by:

a) pga4nla1:

FNL =bnllnpgalow/0.1 8a

b) a1pga4nla2:

FNL =bnllnpgalow/0.1+clnpga4nl/a12 +dlnpga4nl/a1

3 8b

c) a2pga4nl:

FNL =bnllnpga4nl/0.1 8c

where a1 =0.03 g anda2 =0.09 g are assigned threshold levels for linear and non-linear amplification, respectively,pga

low=0.06 g is a variable assigned to transitionbetween linear and nonlinear behaviors, andpga4nl is the predicted PGA in g forVref=760 m/ s, as given by Equation 1 withFS= 0and= 0. The three equations for the non-linear portion of the soil response (Equation 8a8c) are required for two reasons: 1) to

prevent the nonlinear amplification from increasing indefinitely aspga4nldecreases and2) to smooth the transition from linear to non-linear behavior. The coefficients c anddinEquation 8b are given by

c=3ybnlx/x2 9

and

d= 2ybnlx/x3, 10

where

x= lna2/a1 11

and

y=bnllna2/pgalow. 12

The nonlinear slope bnlis a function of both period andVS30 as given by:

a) VS30V1:

bnl=b1. 13a

b) V1VS30V2:

bnl=b1b2lnVS30/V2/lnV1/V2+b2. 13b

c) V2VS30Vref:

bnl=b2lnVS30/Vref/lnV2/Vref . 13c



11/40

d) VrefVS30:

bnl= 0.0. 13dwhereV1 =180 m/ s,V2 =300 m/ s, andb1andb2are period-dependent coefficients (andconsequently,bnlis a function of period as well asVS30). These equations are a simplifiedversion of those used by CS05.

DETERMINATION OF COEFFICIENTS

METHODOLOGY

The selected response variables in the NGA database were first corrected to obtain

the equivalent observations for the reference velocity of760 m/s, using Equations 6, 7,

8a8c, 912, and 13a13d and an equation forpga4nldeveloped early in the project,using only data for which RJB80 km andVS30360 m/s (see BA07 for details). We

then regressed the site-corrected observations to Equation 1 to determine FD andFM.Because the observations had all been corrected to the reference condition, we set FS= 0, simplifying the regression. The analyses were performed using the two-stage regres-sion discussed by Joyner and Boore (1993, 1994); the first stage determines the distancedependence (as well as event terms used in the second stage and the intra-event aleatory

variability, ), and the second stage determines the magnitude dependence (and the

inter-event variability, ). All regressions were done period-by-period; there was nosmoothing of the coefficients that were determined by the regression analyses (althoughsome of the constrained coefficients were smoothed). (Our event term is the average of

the ln Yvalues for a given event, adjusted to the reference velocity and a reference dis-tance (760 m/s and1 km, respectively, in our study). This differs from the event termthat is derived in a random effects model. This latter term is more precisely called a

random effect for a given event (e.g., Abrahamson and Youngs 1992). The residualsfrom our Stage 2 regression are equivalent to this alternate meaning of event term.)

Site Amplification

Because corrections for site amplification were made before doing the first-stage andsecond-stage regressions, we discuss the determination of the site amplification coeffi-cients first. The coefficients in the site-response equations were based on the work ofCS05, rather than determined by our regression analysis. The coefficients needed to

evaluate the site-response equations are listed in Tables 3 and 4. Note that for the refer-

ence velocity of760 m/s, FLIN=FNL =FS= 0. Thus the soil amplifications are specifiedrelative to motions that would be recorded on a B/C boundary site condition.

The rationale for pre-specifying the site amplifications is that the NGA database may

be insufficient to determine simultaneously all coefficients for the nonlinear soil equa-tions and the magnitude-distance scaling, due to trade-offs that occur between param-

eters, particularly when soil nonlinearity is introduced. It was therefore deemed prefer-able to hard-wire the soil response based on the best-available empirical analysis in



12/40

the literature, and allow the regression to determine the remaining magnitude and dis-

tance scaling factors. It is recognized that there are implicit trade-offs involved, and thata change in the prescribed soil response equations would lead to a change in the derivedmagnitude and distance scaling. Note, however, that our prescribed soil response terms

Table 3. Period-dependent site-amplification

coefficients

Period blin b1 b2

PGV 0.600 0.500 0.06

PGA 0.360 0.640 0.14

0.010 0.360 0.640 0.14

0.020 0.340 0.630 0.12

0.030 0.330 0.620 0.11

0.050 0.290 0.640 0.11

0.075 0.230 0.640 0.11

0.100 0.250 0.600 0.13

0.150 0.280 0.530 0.18

0.200 0.310 0.520 0.19

0.250 0.390 0.520 0.16

0.300 0.440 0.520 0.14

0.400 0.500 0.510 0.10

0.500 0.600 0.500 0.06

0.750 0.690 0.470 0.00

1.000 0.700 0.440 0.00

1.500 0.720 0.400 0.00

2.000 0.730 0.380 0.00

3.000 0.740 0.340 0.00

4.000 0.750 0.310 0.00

5.000 0.750 0.291 0.00

7.500 0.692 0.247 0.00

10.000 0.650 0.215 0.00

Table 4. Period-independent site-amplification

coefficients

Coefficient Value

a1 0.03 gpga low 0.06 g

a2 0.09 g

V1 180 m/ s

V2 300 m/ s

Vref 760 m/ s



13/40

are similar to those adopted by other NGA developers who used different approaches;thus there appears to be consensus as to the appropriate level for the soil responsefactors.

The details of setting the coefficients for the soil response equations are as follows.

The linear amplification coefficientsblinwere adopted from CS05. As shown in Figure 3,

they are similar to the linear soil coefficients derived by BJF97. CS05 do not providecoefficients for periods beyond5 s. To determine coefficients for longer periods, we ex-trapolated the blin values as shown on Figure 3. As periods get very long 5 s, wewould expect the relative linear site amplification to decrease (and a trend in this direc-tion has been found by some of the other NGA developers). For this reason, we subjec-

tively decided on the linear trend in terms of log period shown in Figure 3 as the basisfor choosing the values for the longer periods.

The nonlinear slope factorbnldepends on VS30 through the equations given above.Our equations define a somewhat simpler relation than that used by CS05. We compare

the two definitions of the coefficient bnl for periods of 0.2 and3.0 s in Figure 4. Thevalues ofbnlat the hinge points VS30 = V1 andVS30= V2 are given by the coefficients b1andb2, respectively, and these are functions of period. We use CS05s values for most

periods. To extend the value ofb1to periods longer than 5 s we fit two quadratic curvesto the CS05 values: one for all of the values and another for values corresponding to

periods greater than0.2 s; the results were similar (see BA07 for a graph). We based ourvalue ofb1 at periods of7.5 s and10 s on the quadratic fit to all of the CS05 values.

Figure 3. Coefficient controlling linear amplification, as function of period. Values used in

equations in this report indicated by the black dots.



14/40

This curve was also used for the value at 5 s, but the results of using the CS05 value at5 s versus our value makes almost no difference in the predicted ground motions for5 s

periods.

We point out a potential confusion in terminology: according to Equation 8c, FNL=0.0 when pga4nl=0.1 g. Does this mean that there is no nonlinear amplification forthis level of rock motion? Not necessarily. The amplification for this value ofpga4nlisgiven entirely by the FLINterm because the motions used by CS05 to derive the linearamplificationsFLINhad an approximate mean log PGA for most site categories close to0.1 g. FNL is not necessarily zero, however, for values ofpga4nlless than and greaterthan0.1 g. So although the amplification at pga4nl=0.1 g is completely determined by

FLIN, the amplification could implicitly include the nonlinear component that applies forvalues ofpga4nlnear0.1 g. CS05 use only Equation 8c to describe the nonlinear am-

plification, and they do not limit the nonlinear response to pga4nl

0.1 g. It is clearfrom Figure 3 of CS05 and the comment on p. 24 of their paper that they consider Equa-tion 8c to be valid forpga4nlfrom0.02 to 0.8 g. This means that the total amplificationFS can be greater than the linear amplification FLIN for small values of pga4nl;their nonlinear amplification continues to increase without bound as pga4nldecreases.We made an important modification to the CS05 procedure to prevent nonlinear ampli-

fication from extending to small values ofpga4nl, by capping the amplifications at a lowvalue ofpga4nl0.03 g. Simply terminating the nonlinear amplification at a fixed valueofpga4nl results in a kink in plots of ground motion vs. distance. For that reason weincluded a transition curve, as given in Equation 8b.

The total amplification for a short 0.2 s and a long 3.0 s period oscillator isshown in Figure 5 as a function ofpga4nlfor a range ofVS30. At short periods the non-

linear term can result in a significant reduction of motions on sites underlain by rela-tively low velocities. At long periods soil nonlinearity is still important, but the net soil

response effect is an amplification, even for large values ofpga4nl. For periods longerthan 0.75 s (see Table 3) there is no nonlinear contribution to the amplification for

VS30300 m/s

It should be noted that the empirical studies on which the soil amplification functions

Figure 4. Comparison of slope that controls nonlinear amplification function.



15/40

were based contained very few data for hard sites, with VS301,000 m/s. The amplifi-cation functions are probably reasonable for values ofVS30 up to about 1,300 m/s, butshould not be applied for very hard rock sites VS301500 m/s.

Distance Dependence (Stage 1)

The distance dependence of ground motion is determined in the first-stage regres-sion, where the dependent response variable is PGA, PGV, or PSA at a selected period,

in each case corrected to the reference velocity of760 m /s by subtractingFSas definedin Equations 6, 7, 8a8c, 912, and 13a13d from lnYobserved. The corrected responsevariables for our selected subset of the NGA data set (using the exclusion criteria dis-

cussed earlier), with distances out to400 km, are regressed against distance using Equa-tion 14, which is the same as Equation 2 but with dummy variablesc0event added torepresent the event term for each earthquake.

FDRJB,M=c0event+c1+c2M MreflnR/Rref+c3RRref 14

In this equation, c0event is shorthand for the sum

c011+c022+ +c0NENE, 15wherec0jis the event term for event j, jequals 1 for event jand zero otherwise, and

NEis the number of earthquakes.

There are several significant issues in performing this regression. One is that re-gional differences in attenuation are known to exist (e.g., Boore 1989, Benz et al. 1997),

Figure 5. Combined amplification forT=0.2 s andT=3.0 s as function ofpga4nl, for suite of

VS30. Note at short periods (left graph), purely linear amplification does not occur on soft soils

until pga4nl0.03 g.



16/40

even within relatively small regions such as California (e.g., Bakun and Joyner 1984,Boatwright et al. 2003, Hutton and Boore 1987, Mori and Helmberger 1996). We ignorethis potential pitfall and assume that the distance part of the GMPEs apply for crustal

earthquakes in all active tectonic regimes represented by the NGA database. This is a

reasonable initial approach, as the significance of regional effects can be tested later byexamining residual trends (model errors) for subsets of data organized by region. Thesecond difficulty is more problematic: the data in the NGA Flatfile become increasingly

sparse for distances beyond about 80 to 100 km, especially for moderate events. This

makes it difficult, if not impossible, to obtain a robust simultaneous determination ofc1andc3 (slope and curvature). To overcome this database limitation, we have used addi-tional ground-motion data from California that are not in the NGA Flatfile, to first define

the anelastic term,c3, as a function of period. We then used these fixed values ofc3inthe regression of the NGA data set in order to determine the remaining coefficients.

Determination of c3 (anelastic term): The data used to determinec3includes the datacompiled in the NGA database for three small California events, plus many more datafor these same events recorded by accelerometers at broadband stations in California;

these additional data, compiled by J. Boatwright and L. Seekins, were not available fromthe traditional strong-motion data agencies used in compiling the NGA Flatfile. We also

used response variables that we computed from 74 two-component recordings of the

2004 Parkfield mainshock (M 6.0) in the determination ofc3; these data were recordedafter the compilation of the NGA database had concluded. The numbers of stations pro-viding data for our analysis and the corresponding numbers of stations in the NGA Flat-file are given in Table 5 (see also Appendices M and N in BA07).

For the additional data for the three small California earthquakes, we used site

classes assigned by Boatwright and Seekins to correct the response spectra to VS30=760 m/ s. For the Parkfield recordings we did not correct to a common value ofVS30,as we were interested only in determining the distance function, and also because mea-

sured values ofVS30 were available at only a few sites. For all of the data from the four

events we used spectra from the two horizontal components as if they were separate re-cordings (we did not combine the horizontal components). We did the regressions on

this data subset withc1fixed at0.5,0.8, and1.0. We setc2to zero and solved forc3andh. In other words, we fixed a single straight-line slope c1, and then determined thecurvature, c3, required to match the more rapid decay of the data at greater distances (c3must be less than 0). We also solved for the near-source effective depth coefficient, h,

Table 5. Comparisons of numbers of stations in NGA Flatfile and in extended data set used to

determine anelastic coefficient

Earthquake # of Stations in NGA # of Stations Used by BA

2001 AnzaM 4.92 73 1972002 Yorba LindaM 4.27 12 2072003 Big Bear CityM 4.92 37 2622004 ParkfieldM 6.0 0 74



17/40

required to match the less rapid increase of the data as distance decreases at close dis-

tances. An event term that gives the relative amplitude level, c0, is determined for eachof the four earthquakes (these are the coefficients of the dummy variables for eachevent). Figure 6 compares the regression fits to the observations, where the observations

have been normalized to a common amplitude level by subtracting the event termsc0.We chose the c3 values determined for the case c1 =0.8 as the fixedc3 values to apply

in the regression of the NGA data set because c1 =0.8 is a typical value determined inempirical regressions for the effective geometric slope parameter at intermediate periods(BJF97; this study).

As a broader check on the results from our four-event attenuation data set, we de-

termined the best values ofc3 andh to fit the distance functions determined in southernCalifornia from a much larger database, by Raoof et al. (1999). The equivalent values of

c3 andh implied by the Raoof et al. (1999) attenuation results are similar to those thatwe determined from our four-event analysis.

To assign values ofc3over the full period range required in the NGA project, we fita quadratic to the c3 values from the analysis of our four-event data subset. We did notallow the value ofc3at short periods to be less than that for PGA, thus placing an upperlimit on c3 at c3 =0.01151. Similarly, we fixed the values for long periods to be that

determined forT= 3 s, thus placing a lower limit on c3 of c3 =0.00191 (we did notthink it physically plausible for the anelastic attenuation to increase with period at T

3 s).

We also constrained the c3 values for the PGV regressions to be that for the T=1.0 s regression. This choice is a compromise between the similarity in larger-

Figure 6. Normalized ground motions for four events, using extended data set (more data than

in NGA Flatfile). Black curve is regression fit obtained with constraints c 1 =0.8 andc 2 =0.0.



18/40

magnitude scaling that we observed between PGV and PSA at 2 s and the recommen-dation of Bommer and Alarcn (2006) that PGV is related to PSA at 0.5 s.

Determination ofh: It is desirable to constrain the pseudo-depth h in the regressionin order to avoid overlap in the curves for large earthquakes at very close distances. We

did this by performing initial regressions with h as a free parameter, then modifying theresultant values ofh as required to avoid overlap in the spectra at close distances (for thereference site condition of760 m/s). In this regression, c1 was a free variable andc3was constrained to a set of initially-determined values. We fit the values determined with

h as a free parameter with a quadratic, but we observed that the h value at 0.05 s fromthe quadratic fit was very small, much below that determined for PGA. We increased the

h value at 0.05 s to match the value for a regression of PGA with h unconstrained, andrefit the quadratic with this change in the data points. We used the modified quadratic as

the basis for assigning h for all periods. The value ofh at short periods was guided bythe unequivocal statement that PSA is equal to PGA at periods much less than 0.1 s. For

PGA, we adopted the value implied by the modified quadratic for the T= 0.05 s oscilla-tor. We then assigned values ofh for periods between 0.01 s and0.05 s to be the sameas that for0.05 s. Consistent with the convention adopted for the c3coefficient, we usedthe value ofh at 1 s for PGV.

These analyses established smooth, constrained values forc3 andh that facilitatedrobust and well behaved determinations of the remaining parameters by regression of the

NGA database.

Determination of c1, c2, and : With h and c3 constrained, we regressed the re-sponse variables of the NGA database to solve forc1andc2(Equation 3), along with theevent termsc0 for each earthquake, using all data (subject to the exclusions discussedearlier) for distances less than 400 km. The c1 coefficient is the effective geometric

spreading rate (slope) for an event ofM = Mref, while thec2coefficient provides a meansto describe magnitude-dependent distance decay (it changes the slope for events that aregreater or smaller thanMref). The intra-event aleatory uncertainty is given by the stan-dard deviation of the residuals from the Stage 1 regression.

The regression used assigned values for the reference distance, Rref, at which near-source predictions are pegged, and for the reference magnitude, Mref, to which the mag-nitude dependence of the geometric spreading is referenced. The assigned values for

these reference values are arbitrary and are largely a matter of convenience. ForMref, wechose a value of 4.5, since this is the approximate magnitude of much of the data used

to determine the fixedc3 coefficients; this choice means that the magnitude dependenceof the slope will be referenced to that observed for small events. ForRref, we use thevalue of1 km. This is convenient because the curves describing the distance dependence

pivot aroundR =Rref. The curves for larger magnitudes are flatter than for smaller mag-nitudes, which can lead to those curves being below the curves for smaller magnitudes at

distances less than the pivot distance. This was avoided by choosing Rref=1 km, al-though any value such that Rrefminh, where the minimum is taken over all periods,



19/40

would prevent undesirable overlapping of prediction curves near the source (i.e. we want

to ensure that R will always be greater than the pivot distance ofRref, even when RJB

=0 km).

Magnitude Dependence (Stage 2)

The event terms (coefficientsc0jin Equation 14 from the Stage 1 regression wereused in a weighted Stage 2 regression to determine the magnitude scaling of the re-sponse variables. As discussed in Joyner and Boore (1993), the Stage 2 weighted regres-

sion was iterative in order to solve for the inter-event variability . The basic form we

selected for the magnitude scaling is a quadratic, similar to the form used by BJF97.

However, we imposed a constraint that the quadratic not reach its maximum at M

8.5, in order to prevent oversaturation (the prediction of decreasing amplitudes withincreasing magnitude). The following algorithm was used to implement the constrainedquadratic magnitude dependence:

1. Fit the event termsc0j for a given period to a second-order polynomial. If theM for which the quadratic starts to decrease Mmax is greater than 8.5, weadopt this regression for the magnitude dependence for this period.

2. If Mmax for a given period is less than 8.5, we perform a two-segment regres-

sion, hinged atMh(described below), with a quadratic forMMhand a linear

function forMhM. If the slope of the linear function is positive, we adopt thistwo-segment regression for the magnitude dependence for this period.

3. If the slope of the linear segment is negative, we redo the two-segment regres-sion for that period, constraining the slope of the line above Mh to be 0.0. Note

that the equations for almost all periods less than or equal to 1.0 s required theconstraint of zero slope; this is telling us that for short periods the data actuallyindicate oversaturation. We felt that because of limited data and knowledge,

oversaturation was too extreme at this stage of equation development, and wechose to impose saturation rather than allow the data to dictate an oversaturated

form. More observations from ground motions near large earthquakes, as wellas theoretical simulations using dynamic rupture models (e.g., Schmedes andArchuleta, 2007) may give us confidence in allowing oversaturation in futureversions of GMPEs.

Choice of Mh: The parameter Mh is the hinge magnitude at which the constrainedmagnitude scaling in the two-segment regression changes from the quadratic form to thelinear form. Subjective inspection of nonparametric plots of data clearly indicated thatnear-source ground motions at short periods do not get significantly larger with increas-

ing magnitude, beyond a magnitude in the range of 6.5 to 7, and therefore we set Mhwithin this range.

Fault-Type Dependence: Plots of event terms against magnitude (presented later)showed that normal-fault earthquakes have amplitudes that are consistently below those

for strike-slip and reverse earthquakes for most periods (others have found similar re-sults, including Spudich et al. 1999, Bommer et al. 2003, and Ambraseys et al. 2005).We used this observation to guide our determination of the dependence on fault type. We

first grouped the data from all fault types together and solved for the coefficients e1,e5,



20/40

e6, e7, ande8 in Equation 5a and 5b), setting e2, e3, ande4 to 0.0. The regression wasthen repeated, fixing the coefficientse5,e6,e7, ande8to the values obtained when lump-

ing all fault types together, and solving for the coefficients e2,e3, ande4of the fault typedummy variables SS, NS, and RS. Thus we have constrained the relative scaling of am-

plitudes with magnitude to be the same for all event types, but we allow an offset in theaverage predicted amplitude level according to the fault mechanism. The inter-event

aleatory uncertainty was slightly different for these two cases, so subscripts U andM were used to distinguish between unspecified and specified fault type, respectively,in the table of aleatory uncertainties. Note that the term unspecified is strictly appli-cable to a random selection of an earthquake from the distribution of fault types used in

our analysis; it is an accurate description of a truly random selection from all earth-quakes only to the extent that the distribution of all fault types is equal to the distributionused in our analysis.

All analyses were done using Fortran programs developed by the first author, in some

cases incorporating legacy code from programs and subroutines written by W. B. Joyner.

RESULTS

COEFFICIENTS OF THE EQUATIONS

The coefficients for the GMPEs are given in Tables 3, 4, and 68. The coefficients

are for lnY, where Yhas units of g for PSA and PGA and cm/s for PGV. The units ofdistance and velocity are km and m/s, respectively. The equation forpga4nlis the sameas for PGA, with VS30760 m/s (for which FS= 0) (Boore and Atkinson, 2008).

There are no normal-fault data are in our data set for an oscillator period of10 s, andthus formally we could not obtain the coefficient e3 for that period; the value in Table 7was obtained using the assumption that the ratio of motions for normal and unspecified

faults is the same for periods of 7.5 s and 10 s. With this assumption, e310s= e110s + e37.5s e17.5s.

Fit of the Stage 1 Regressions

BA07 contains a series of graphs showing the observations in comparison to the

Stage 1 regression predictions. These figures provide a visual test of the ability of ourfunctional form to represent the distance dependence of the response variables. A more

precise way of looking for systematic mismatches between predictions and observationsis to plot the residuals from the Stage 1 analysis, defined as the ratio of observed to

predicted ground motions. Figure 7 shows residuals as a function of distance for PGA

and for PSA at 10 s; these span the range of seismic intensity measures included in ourequations (the graphs of residuals for PSA at most oscillator periods are similar to that

for PGA in Figure 7see BA07 for a complete set of graphs). For the sake of clarity, wehave separated the residuals into different magnitude ranges and for two specific earth-quakes in Figure 7. Log residuals averaged over distance bins 0.1 log unit in width and

magnitude bins1 unit in width are shown in Figure 8 for two representative periods; thegraphs are grouped by values ofVS30. While there are some small departures from a nullresidual (values of 1 and 0 in Figures 7 and 8, respectively), there are no significant



21/40

trends in magnitude, distance, or shear-wave velocity. We therefore judge the fit betweenobservations and our predictions to be reasonable. In particular, we note that the im-

posed soil response coefficients appear to be adequate, as evidenced by the apparent fitover the three distinct ranges of shear-wave velocity used in Figure 8; the fit is good at

both short and large distances over all magnitude ranges, which implicitly supports the

degree of nonlinearity that was specified.

Fit of the Stage 2 Regressions

Figure 9 is a plot of the antilog of the event terms c0jfrom the Stage 1 regression

as a function of magnitude, with the Stage 2 regression fit to these terms superimposed.The fault type for each earthquake is indicated, as are curves for fault type unspecified

and for strike-slip, normal, and thrust/reverse faults (the fault type is indicated by thecolor of the symbols). The functional form provides a reasonable fit to the near-source

Table 6. Distance-scaling coefficients (Mref=4.5 and

Rref=1.0 km for all periods, except Rref=5.0 km for

pga4nl)

Period c1 c2 c3 h

PGV 0.87370 0.10060 0.00334 2.54

PGA 0.66050 0.11970 0.01151 1.35

0.010 0.66220 0.12000 0.01151 1.35

0.020 0.66600 0.12280 0.01151 1.35

0.030 0.69010 0.12830 0.01151 1.35

0.050 0.71700 0.13170 0.01151 1.35

0.075 0.72050 0.12370 0.01151 1.55

0.100 0.70810 0.11170 0.01151 1.68

0.150 0.69610 0.09884 0.01113 1.86

0.200 0.58300 0.04273 0.00952 1.98

0.250 0.57260 0.02977 0.00837 2.07

0.300 0.55430 0.01955 0.00750 2.14

0.400 0.64430 0.04394 0.00626 2.24

0.500 0.69140 0.06080 0.00540 2.32

0.750 0.74080 0.07518 0.00409 2.46

1.000 0.81830 0.10270 0.00334 2.54

1.500 0.83030 0.09793 0.00255 2.66

2.000 0.82850 0.09432 0.00217 2.73

3.000 0.78440 0.07282 0.00191 2.83

4.000 0.68540 0.03758 0.00191 2.89

5.000 0.50960 0.02391 0.00191 2.93

7.500 0.37240 0.06568 0.00191 3.00

10.000 0.09824 0.13800 0.00191 3.04



22/40

amplitude data. Note that the magnitude scaling forT= 10 s at M6.5 is strongly con-trolled by the data from only one small earthquake (2000 Yountville, M 5.0), and may

therefore be unreliable forM6.5.

Predictions of PSA from Combined Stage 1 and Stage 2 Regressions

Graphs of PSA predicted from our equations for three values ofRJB and four mag-nitudes are shown in Figure 10. The curves for the larger earthquakes tend to squeeze

together for periods near0.20.3 s, probably a reflection of the pinching together of theeffective geometric spreading factor for these periods. But otherwise the PSA are quitesmooth, especially considering that many of the coefficients were determined indepen-dently for each period.

Plots of PSA as a function of distance are shown in Figure 11 for two representative

periods (see BA07 for plots at other periods). The figure is forVS30 =760 m/ s (NEHRPB/C boundary).

The effect of VS30 on predicted ground-motion amplitude is shown in Figure 12.Nonlinear soil amplification causes the curves to cross, such that at close distances lower

Table 7. Magnitude-scaling coefficients

Period e1 e2 e3 e4 e5 e6 e7 Mh

PGV 5.00121 5.04727 4.63188 5.08210 0.18322 0.12736 0.00000 8.50

PGA 0.53804 0.50350 0.75472 0.50970 0.28805 0.10164 0.00000 6.75

0.010 0.52883 0.49429 0.74551 0.49966 0.28897 0.10019 0.00000 6.75

0.020 0.52192 0.48508 0.73906 0.48895 0.25144 0.11006 0.00000 6.75

0.030 0.45285 0.41831 0.66722 0.42229 0.17976 0.12858 0.00000 6.75

0.050 0.28476 0.25022 0.48462 0.26092 0.06369 0.15752 0.00000 6.75

0.075 0.00767 0.04912 0.20578 0.02706 0.01170 0.17051 0.00000 6.75

0.100 0.20109 0.23102 0.03058 0.22193 0.04697 0.15948 0.00000 6.75

0.150 0.46128 0.48661 0.30185 0.49328 0.17990 0.14539 0.00000 6.75

0.200 0.57180 0.59253 0.40860 0.61472 0.52729 0.12964 0.00102 6.75

0.250 0.51884 0.53496 0.33880 0.57747 0.60880 0.13843 0.08607 6.75

0.300 0.43825 0.44516 0.25356 0.51990 0.64472 0.15694 0.10601 6.75

0.400 0.39220 0.40602 0.21398 0.46080 0.78610 0.07843 0.02262 6.75

0.500 0.18957 0.19878 0.00967 0.26337 0.76837 0.09054 0.00000 6.75

0.750 0.21338 0.19496 0.49176 0.10813 0.75179 0.14053 0.10302 6.75

1.000 0.46896 0.43443 0.78465 0.39330 0.67880 0.18257 0.05393 6.75

1.500 0.86271 0.79593 1.20902 0.88085 0.70689 0.25950 0.19082 6.75

2.000 1.22652 1.15514 1.57697 1.27669 0.77989 0.29657 0.29888 6.75

3.000 1.82979 1.74690 2.22584 1.91814 0.77966 0.45384 0.67466 6.75

4.000 2.24656 2.15906 2.58228 2.38168 1.24961 0.35874 0.79508 6.75

5.000 1.28408 1.21270 1.50904 1.41093 0.14271 0.39006 0.00000 8.50

7.500 1.43145 1.31632 1.81022 1.59217 0.52407 0.37578 0.00000 8.50

10.000 2.15446 2.16137 2.53323 2.14635 0.40387 0.48492 0.00000 8.50



23/40

values ofVS30(softer sites) will have lower predicted amplitudes than stiffer sites, due tononlinear deamplification. The effect is more pronounced at short periods than at long

periods.

Surface Slip vs. No-Surface Slip Earthquakes

Several authors (e.g., Somerville and Pitarka 2006) have proposed that the high-frequency ground motions from earthquakes with faults that break to the surface aresmaller than from those with faults that remain buried. We search for evidence of this

effect in Figure 13, which shows the event-term residuals from the Stage 1 regression

plotted againstM for the two classes of earthquakes. The first thing to notice is that mostsurface-slip earthquakes correspond to larger magnitudes, with almost no buried rup-

tures for magnitude greater than M = 7. For this reason any reduction in motions forsurface-slip earthquakes will be mapped into reduced magnitude scaling in the Stage 2magnitude regression. In order to differentiate magnitude scaling from the effects of sur-face versus buried rupture, data from both class of rupture are needed for the same range

Table 8. Aleatory uncertainties (: intra-event uncertainty; : inter-event

uncertainty; T: combined uncertainty2 + 2; subscriptsU,Mfor faulttype unspecified and specified, respectively)

Period U TU M TM

PGV 0.500 0.286 0.576 0.256 0.560

PGA 0.502 0.265 0.566 0.260 0.564

0.010 0.502 0.267 0.569 0.262 0.566

0.020 0.502 0.267 0.569 0.262 0.566

0.030 0.507 0.276 0.578 0.274 0.576

0.050 0.516 0.286 0.589 0.286 0.589

0.075 0.513 0.322 0.606 0.320 0.606

0.100 0.520 0.313 0.608 0.318 0.608

0.150 0.518 0.288 0.592 0.290 0.594

0.200 0.523 0.283 0.596 0.288 0.596

0.250 0.527 0.267 0.592 0.267 0.592

0.300 0.546 0.272 0.608 0.269 0.608

0.400 0.541 0.267 0.603 0.267 0.603

0.500 0.555 0.265 0.615 0.265 0.615

0.750 0.571 0.311 0.649 0.299 0.645

1.000 0.573 0.318 0.654 0.302 0.647

1.500 0.566 0.382 0.684 0.373 0.679

2.000 0.580 0.398 0.702 0.389 0.700

3.000 0.566 0.410 0.700 0.401 0.695

4.000 0.583 0.394 0.702 0.385 0.698

5.000 0.601 0.414 0.730 0.437 0.744

7.500 0.626 0.465 0.781 0.477 0.787

10.000 0.645 0.355 0.735 0.477 0.801



24/40

Figure 7. Stage 1 residuals, separated by magnitude, for PGA and T=10 s PSA (these span the

range of possibilities; see BA07 for more plots). The residuals for the 1999 Chi-Chi earthquake

are shown separately, in the bottom two graphs.



25/40

of magnitudes. As seen in Figure 13, it is only for strike-slip earthquakes that there is

more than one of each class of earthquake in a common magnitude range (there are sev-eral strike-slip events of 5.76.7 in both classes). There is no indication for these earth-quakes that the event-term residuals are systematically different for the two classes ofdata. Therefore, there was no need to include dummy variables for surface-slip/buriedearthquakes in our functional forms. As confidence in simulations from dynamic modelsof rupture propagation increases, or if additional data change our understanding, it might

be that in the future we will add a buried/surface faulting term to the equations. By do-ing so, the apparent saturation of the magnitude scaling would not be as dramatic (i.e.,the larger earthquakes are entirely surface slip events, and if these produce smallerground motions than buried events, as has been suggested by Somerville and colleagues

Figure 8. Stage 1 regression residuals log10 units for 0.2-s and 3-s response spectra, aver-aged over distance bins 0.1 log unit in width and magnitude bins 1 unit in width. Only bins with

at least three observations are plotted. The standard error of the mean is shown for the middlemagnitude bin (6 to 7) only. Two ranges of shear-wave velocity are shown: 180 360 m / s (top),

360760 m/s (bottom). No residuals are shown forVS30760 m / s because of the small num-

ber of observations in that velocity range.



26/40

Figure 9. Yevent (the antilog of the average of the log of the motions for each event adjustedto the reference distance of 1 km and the reference velocity of 760 m/s; in the terminology of

Equations 14 and 15, Yevent =expc0event) and Stage 2 regression fits. Note that the samevertical scale was used for all graphs in order to compare the magnitude scaling from period to

period.



27/40

(e.g., Somerville and Pitarka 2006), then there will be an apparent tendency for satura-tion if the events are not separated into two classes according to whether they break tothe surface or not).

Dependence of Stage 1 Residuals on Basin Depth

Another ground-motion effect that we searched for in the residuals of the Stage 1regression was that of basin depth. Basin-depth effects on ground-motion amplitudeshave been reported in empirical studies (Field, 2000; Choi et al. 2005), and from simu-lations (Day et al. 2008). One of the reasons that we did not include a basin-depth term

in our equations is indicated in Figure 14, which shows the distribution ofVS30 and a

Figure 10. PSA from our equations, as function of period. The spectra are shown for three

distances and four magnitudes, for fault type unspecified andVS30 =760 m/s.



28/40

Figure 11. PSA from our equations, as a function of distance. The spectra are shown for mag-

nitudes 5, 6, 7, and 8, for fault type unspecified andVS30 = 760 m / s. Note that the same vertical

scale was used for both graphs in order to compare the magnitude scaling for the two periods.

Figure 12. PSA from our equations, as a function of distance for magnitude 7, fault type un-

specified, and three values ofVS30. Note that the same vertical scale was used for both graphs

in order to compare the magnitude scaling for the two periods.



29/40

measure of basin depth. The plot shows all data in the NGA Flatfile for which both VS30have been measured and basin depth have been estimated. It is clear that the softer sites

are in basins, and hence basin depth and VS30 are strongly correlated (this was foundpreviously by Choi et al. 2005, Figure 7). Therefore any basin depth effect will tend to

Figure 13. Antilogarithms of Stage 2 residuals, plotted against magnitude and differentiated by

events of different fault types, for which faults did or did not break to surface.

Figure 14. VS30plotted against one measures of basin depth: the depth to a shear-wave velocity

of 1.5 km/s. All values in the NGA Flatfile with basin depths and measured (rather than esti-

mated) values ofVS30 are shown.



30/40

have been captured by the empirically-determined site amplification. To try to separatethe amplification and the basin-depth effects in the data would require use of additionalinformation or assumptions. Since we are opting for the simplest equations required by

the data, no attempt was made to break down the site-response function into basin depthand the amplification terms.

We searched for any uncaptured basin depth effect by examining the residuals of theStage 1 regressions. We find that the residuals have no dependence on basin depth, ex-cept for a trend to positive residuals (underprediction) for long periods at distances be-

yond80 km(by about a factor of 1.6), for sites having depth-to- 1.5 km/s700 m. (Thetrend for greater distances is shown in a figure not included here.) These residuals couldalso be due to regional variations in the distance function, along with correlations be-tween distance and the basin depth. Figure 15 contains plots of the Stage 1 residuals

against the depth-to-VS30=1.5 km/s; only residuals forRJB80 kmare shown, in ordernot to map mismatches in the more distant attenuation into the residuals. There is no

obvious dependence of the residuals on basin depth. But assuming that the positive re-

siduals at distances greater than 80 kmare due solely to a basin depth effect, the trendsindicate that our equations may underpredict long-period motions at large distancesfrom sites in deep basins. For shallower basins and at shorter distances, we find no basindepth effect. This is not surprising in light of the observations made above regarding the

correlation between basin parameters and VS30. (Note: similar results were obtainedwhen the depth to2.5 km/s was used as the measure of basin depth.) Another reason for

Figure 15. Stage 1 residuals plotted against depth to VS= 1.5 km/ s, differentiated byVS30, for

RJB80 km.



31/40

little apparent basin effect has to do with the way in which basins affect incoming waves;Choi et al. (2005) found a basin effect for sources inside the basin in which the motions

were recorded, but little effect for sources outside the basin; they attribute the differenceto the manner in which incoming waves are converted and refracted upon entering the

basin. As many of our data come from earthquakes that occurred outside the basins in

which they were recorded, a similar explanation might apply to our finding.

Comparison of GMPEs Developed With and Without the 1999

Chi-Chi Earthquake

Because the Chi-Chi earthquake forms a significant fraction of the data set we used

in developing our equations, it is important to see how the equations would change if thedata from the Chi-Chi earthquake were eliminated from both the Stage 1 and the Stage2 regressions. We therefore repeated the complete analysis without the Chi-Chi data.Figure 16 compares selected ground-motion intensity measures given by the two sets of

equations. The figure also shows the percent of data used in the regression analysis fromthe Chi-Chi earthquake (the number of Chi-Chi recordings is the numerator of the ratio).It is clear that the fraction of the data set contributed by the Chi-Chi earthquake in-

creases with period, reaching 64% of the data set for a period of10 s. For this reason itis not surprising that the predictions of10 s PSA are quite different for the equationsdeveloped with and without the Chi-Chi data (the ordinate scales of all graphs in Figure

16 are the same, to facilitate comparisons of the relations between the two predictionsbetween periods); at intermediate to short periods, however, the differences are not dra-matic. Interestingly, the differences can occur even at small magnitudes (despite the factthat we include only the Chi-Chi mainshock, not its aftershocks). We think the explana-tion of this apparent paradox is that the Chi-Chi earthquake is very well recorded and

thus dominates the Stage 1 regression, for which each recording of an earthquake hasequal weight in determining the distance terms in the equations. These distance termsthen affect the event terms, and this in turn controls the magnitude scaling. We concludethat although the Chi-Chi earthquake affects the GMPEs, it is only a major controlling

factor in the predictions of PSA at periods of greater than 5 s.

Comparison of BA07 and BJF97 GMPEs

It is interesting to compare our new predicted ground motions with those from theBoore et al. (1997) (BJF97) equations. Figure 17 (top row) compares the magnitude-distance distribution of the data used in each study. It is apparent that many more data

are used in the new equations; the NGA data fill gaps at close distances for all magni-tudes, add more data at small magnitudes at all distances, add data for large magnitudes,and fill out the distribution so that no longer is there a strong correlation between dis-tance and magnitude in the data set. For this reason, the new equations provide a morerobust prediction of ground-motion amplitudes over a wide range of magnitudes and

distances.

We compare predicted ground motions from the BJF97 equations and from our cur-

rent equations in Figure 17 (bottom row), forVS30=420 m/ s, which is near the weightedgeometric mean of the velocities for the sites used in the BJF97 regression analysis. We



32/40

use the same scale for the ordinates in both graphs. The new and old equations predict

similar amplitudes for M andRJB ranges for which data were available for the BJF97equation development. Large differences occur in regions of the magnitude-distancespace for which data were not available in BJF97; the differences in the predicted valuesof seismic ground-motion intensity are largely attributable to the overly-simplified

distance-independent magnitude scaling used in the BJF97 equations.At all periods, the new equations predict significantly smaller motions than do the

BJF97 equations for large magnitudes. This is probably the most important change in thenew equations compared to the old equations. The difference in the predicted motions is

particularly large forT= 1 s andM =7.5(a factor of 2.4 at RJB=1 km). Almost no datawere available in BJF97 forM 7.5 andRJB10 km (see Figure 17), so discrepancies

Figure 16. Comparisons of PSA for four periods from equations developed with and without

1999 Chi-Chi mainshock. Ratios are number of Chi-Chi recordings used to develop final equa-

tions divided by total number of recordings.



33/40

are not surprising. The BJF97 data were forRJBcentered about30 km. The discrepancybetween the predictions from the BJF97 and the new equations is not nearly as strong for

RJBnear30 km as it is forRJB10 km. Observed differences at RJB 30 kmare likelydue to including more data for large earthquakes in our current equations. The values of

the BJF97 motions at close distances are strongly controlled by the assumption ofdistance-independentM scaling (and therefore the scaling at close distances is driven by

the RJB 30 km data). The current equations allow for M-dependent distance scaling.

Another effect that can reduce motions predicted from our equations at close distancesfrom large earthquakes is nonlinear site response, which is not included in BJF97.

The total aleatory uncertainties, as well as the intra- and inter-event uncertainties, aresignificantly larger for the new equations than for the BJF97 equations (e.g., for a period

of0.2 s the total aleatory uncertainty is 0.60 for our equations and 0.44 for BJF97; morecomparisons can be found in Table 4.6 of BA07). We are not sure of the reasons for the

Figure 17. Top two graphs: Comparison of magnitude-distance distribution of data used by

BJF97 and by BA08; bottom two graphs: PGA and T=1.0 s PSA predictions from BJF97 and

from BA08 GMPEs.



34/40

differences, but we suspect the differences are due to a combination of more data pro-viding a better sample of the true uncertainty, as well as an increase in the uncertainties

produced by mixing data from regions for which the attenuation of the motions might bedifferent (see Douglas 2007, for a study of regional differences in ground-motion pre-diction). The larger sigma values will offset to some extent the smaller ground motions

for large magnitudes in the construction of seismic hazard maps. However, it is a pointfor further investigation how the aleatory uncertainties should be implemented in hazardanalyses for a particular site, given that part of the aleatory uncertainty arises from mix-ing data from numerous regions.

GUIDELINES FOR USAGE

LIMITS ON PREDICTOR VARIABLES

We wish to emphasize that our equations should be used only for predictor variables

in these ranges: M = 5 8

RJB200 km

VS30 =1801300 m/ sThese limits are subjective estimates based on the distributions of the recordings used todevelop the equations.

PREDICTIONS FOR OTHER MEASURES OF SEISMIC INTENSITY

The NGA GMPEs are for the GMRotI measure of seismic intensity. Simple conver-sion factors between GMRotI and other measures of seismic intensity are given byBeyer and Bommer (2006) and Watson-Lamprey and Boore (2007), as well as by Camp-

bell and Bozorgnia (2008).

DISCUSSION AND SUMMARY

We have presented a set of ground-motion prediction equations that we believe arethe simplest formulation demanded by the NGA database used for the regressions. Fu-ture versions of the equations might include additional terms, such as basin depth, if

these can be unambiguously supported by data. Expansion of the NGA database by wayof additional or reprocessed data could potentially support the inclusion of more predic-tive variables. In spite of this, we note that the aleatory uncertainties in our equations aresimilar to those of other NGA developers who included more predictive variables.Therefore we do not think that our simplified analysis limits the usefulness of our equa-

tions, at least for those situations for which predictor variables not included in our equa-

tions are not crucial in site-specific hazard analysis.One modification we would like to address in future versions of our equations is to

account for regional variations in distance attenuation, particularly at distances beyond

about80 km. The near-source data could be used to constrain magnitude scaling for allregions, which could be patched onto regionally-dependent distance functions. The ap-

proach taken in this study, in which the anelastic coefficient was constrained using data



35/40

from a few earthquakes in central and southern California, is not optimal. Furthermore,there are inconsistencies in the pseudo-depths that might be attributed to forcing the val-

ues of the anelastic coefficient into the regression of the worldwide data set. Notwith-standing these limitations, the new relations developed here provide a demonstrably re-liable description of recorded ground-motion amplitudes for shallow crustal earthquakes

in active tectonic regions over a wide range of magnitudes and distances.

ACKNOWLEDGMENTS

This study was sponsored by the Pacific Earthquake Engineering Research Centers

Program of Applied Earthquake Engineering Research of Lifelines Systems supportedby the California Department of Transportation, the California Energy Commission, andthe Pacific Gas and Electric Company.

This work made use of the Earthquake Engineering Research Centers Shared Facili-ties supported by the National Science Foundation, under award number EEC-9701568through the Pacific Earthquake Engineering Research (PEER) Center. Any opinions,findings, and conclusions or recommendations expressed in this material are those of the

authors and do not necessarily reflect those of the National Science Foundation.

We have benefited from discussions and comments from many people. First andforemost, we want to thank the whole PEER NGA project team for the opportunity to

participate in the project; all interactions with the members of the team were extraordi-narily open and supportive. In addition, we thank these people, in alphabetical order:Sinan Akkar, Jack Boatwright, John Douglas, Art Frankel, Vladimir Graizer, SteveHarmsen, Robert Herrmann, Tom Holzer, Charles Mueller, Maury Power, Rakesh Sai-

gal, Linda Seekins, and Chris Stephens. John Douglas and Jon Stewart provided excel-lent reviews that substantially improved the paper. Finally, we acknowledge the contri-

butions of Bill Joyner to the development of the regression procedures and codes used inthis work.

APPENDIX A. CHOICE OF VS30FOR A NEHRP CLASS

The need sometimes arises to evaluate GMPEs for a particular NEHRP site class.

Because the PEER NGA GMPEs use the continuous variableVS30as the predictor vari-able for site amplification, the question naturally arises as to what value ofVS30 to usefor a specific NEHRP class. To explore that question, we used the distribution ofVS30values from the borehole compilation given in Boore (2003) and from the NGA Flatfile,

and computed the geometric means of the average of the VS30 values in each NEHRPclass.

We used the geometric mean ofVS30 in each NEHRP class, as these will give the

same value of lnYas the average of the ln Ys obtained using the actual VS30 values inthe data set. Here is the analysis:

Because



36/40

lnYblnV30

the average ofln Yfor a number ofVS30s in a site class is:

lnYb1

Ni=1

N

lnV30i

and the same value ofln Yis obtained using the value ofVS30given by:

lnV30 =1

Ni=1

N

lnV30i

But does that mean that the values ofVS30 in the NGA database should be used todetermine the average value ofVS30that will be substituted into the GMPEs for a given

NEHRP site class? Yes, under the assumption that the distribution ofVS30 in the NGA

database is similar to the one that would be obtained if a random site were selected. Wediscuss this in more detail at the end of this appendix.

To determine the geometric means ofVS30from the NGA Flatfile, we used the Excelfunctionvlookup to select only one entry per station. Figure A1 shows the histograms.

For the Boore (2003) data set, we used values ofVS30 for which the borehole velocitieshad to be extrapolated less than 2.5 m to reach 30 m. The top graph shows histogramsfor the Boore (2003) velocities; the middle graph shows histograms for NGA velocities

for which the values ofVS30are based on measurements (source=0 and 5); and the bot-tom graph is for NGA values from measurements and estimations (source=0, 1, 2, and5). In choosing the most representative value ofVS30 for each NEHRP class, we gavemost weight to the middle graph in Figure A1. Those histograms used more data than inBoore (2003), but they are not subject to the possible bias in using an estimated value of

VS30, in which the value might be based on the assignment of a NEHRP class to a site,with someone elses correlation between NEHRP class andVS30 (correlations that mayor may not have used the geometric mean ofVS30). We are trying to find the appropriatevalue independently.

The gray vertical lines in Figure A1 are the geometric means in each NEHRP class

for the data used for each graph; the black vertical lines in Figure A1 are the VS30valueswe recommend be used for each NEHRP class; they are controlled largely by the analy-

sis of the source=0 and 5 NGA data. Table A1 contains the values ofVS30 determinedfor the different histograms. Based on these values, the second-to-last column in thetable contains the observation-based representative values that could substituted into the

NGA GMPEs for specific NEHRP classes. The last column contains another possible setof values for evaluating the GMPEs for a specific NEHRP class; these values are thegeometric means of the velocities defining each NEHRP class, rounded to the nearest

5 m/ s (e.g., for NEHRP class D the value from the class definition is180360=255 m/ s).

As mentioned before, the values in the second-to-last column of Table A1 are validrepresentations of the different NEHRP classes if the distribution of velocities in the



37/40

Figure A1. Histograms used to determine value ofVS30 to use in evaluating NGA GMPEs for

a particular NEHRP class (gray vertical lines are the geometrical means of the VS30 in each

NEHRP site class, and black vertical lines are recommended values for each NEHRP class, as

given in the second-to-last column in Table A1).



38/40

geographic region of interest is the same as that for the data used in the analysis above.

Most of the measured values in the NGA database, however, come from the Los Angeles

and San Francisco areas of California, so there is the potential for a bias if the VS30val-ues for those regions are not representative of a generic site. An alternative set of rep-

resentative VS30values for each NEHRP site class is given by the geometric mean of thevelocities defining the site-class boundaries. These are given in the last column of TableA1. The values in the last two columns of Table A1 are similar, but to assess the impactof the two sets of representative values, we evaluated the ratios of ground motions forthe two values for each NEHRP class, for a wide range of periods and distances. The

differences in ground motions using the two possible sets ofVS30 values are less than8%, 5%, and 3% for NEHRP classes B, C, and D, respectively. The differences are larg-est at long periods for classes B and C and for short periods for class D. The differences

in ground motions for each site class obtained using the alternative sets of representativeVS30 values are so small that either set of could be used.

REFERENCES

Abrahamson, N. A., and Youngs, R. R., 1992. A stable algorithm for regression analysis using

the random effects model, Bull. Seismol. Soc. Am. 82, 505510.

Ambraseys, N. N., Douglas, J., Sarma, S. K., and Smit, P. M., 2005. Equations for the estima-

tion of strong ground motions from shallow crustal earthquakes using data from Europe and

the Middle East: Horizontal peak ground acceleration and spectral acceleration, Bull. Earth-

quake Eng. 3, 153.

Atkinson, G. M., 1993. Earthquake source spectra in eastern North America,Bull. Seismol. Soc.

Am. 83, 17781798.

Bakun, W. H., and Joyner, W. B., 1984. The ML scale in central California, Bull. Seismol. Soc.

Am. 74, 18271843.

Benz, H. M., Frankel, A., and Boore, D. M., 1997. Regional Lgattenuation for the continental

United States, Bull. Seismol. Soc. Am. 87, 606619.

Beyer, K., and Bommer, J. J., 2006. Relationships between median values and between aleatory

Table A1. Correspondence between NEHRP class and geometric mean VS30 (see text)

NEHRP

Site

Class

GeometricMean of

Measured

VS30 in

NGA

Flatfile

GeometricMean of

Measured &

InferredVS30in NGA

Flatfile

Geometric

Mean of

VS30 in

Boore (2003)

Suggested

VS30 Based

on Measured

Velocities in

boore_atkinson_2008_ground motion prediction equations for the average horizontal component of pga,...

Documents